How to interpret the GPU utilization in Nsight Systems profile?


I’m using the Nsight Systems profiling tool to analyze the GPU utilization of my process.
Here is one of profile I get:

  • In the profile, I’m seeing the full sky-blue bars in the CUDA HW row, does it mean that ALL GPU compute resources (i.e., all SMs in the GPU) are fully utilized during that time? Or else, is it showing the compute resource utilization of a single SM running at the time?
  • If it’s showing a single SM, how do I know which SM it is showing? I see there are two active streams (Stream 7 and 13), and both of them are active in parallel for some durations.

Any comments are much appreciated!